arch/arm64/kvm/hyp/nvhe/mem_protect.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2020 Google LLC
   4  * Author: Quentin Perret <qperret@google.com>
   5  */
   6
   7 #include <linux/kvm_host.h>
   8 #include <asm/kvm_emulate.h>
   9 #include <asm/kvm_hyp.h>
  10 #include <asm/kvm_mmu.h>
  11 #include <asm/kvm_pgtable.h>
  12 #include <asm/stage2_pgtable.h>
  13
  14 #include <hyp/switch.h>
  15
  16 #include <nvhe/gfp.h>
  17 #include <nvhe/memory.h>
  18 #include <nvhe/mem_protect.h>
  19 #include <nvhe/mm.h>
  20
  21 #define KVM_HOST_S2_FLAGS (KVM_PGTABLE_S2_NOFWB | KVM_PGTABLE_S2_IDMAP)
  22
  23 extern unsigned long hyp_nr_cpus;
  24 struct host_kvm host_kvm;
  25
  26 struct hyp_pool host_s2_mem;
  27 struct hyp_pool host_s2_dev;
  28
  29 /*
  30  * Copies of the host's CPU features registers holding sanitized values.
  31  */
  32 u64 id_aa64mmfr0_el1_sys_val;
  33 u64 id_aa64mmfr1_el1_sys_val;
  34
  35 static const u8 pkvm_hyp_id = 1;
  36
  37 static void *host_s2_zalloc_pages_exact(size_t size)
  38 {
  39         return hyp_alloc_pages(&host_s2_mem, get_order(size));
  40 }
  41
  42 static void *host_s2_zalloc_page(void *pool)
  43 {
  44         return hyp_alloc_pages(pool, 0);
  45 }
  46
  47 static int prepare_s2_pools(void *mem_pgt_pool, void *dev_pgt_pool)
  48 {
  49         unsigned long nr_pages, pfn;
  50         int ret;
  51
  52         pfn = hyp_virt_to_pfn(mem_pgt_pool);
  53         nr_pages = host_s2_mem_pgtable_pages();
  54         ret = hyp_pool_init(&host_s2_mem, pfn, nr_pages, 0);
  55         if (ret)
  56                 return ret;
  57
  58         pfn = hyp_virt_to_pfn(dev_pgt_pool);
  59         nr_pages = host_s2_dev_pgtable_pages();
  60         ret = hyp_pool_init(&host_s2_dev, pfn, nr_pages, 0);
  61         if (ret)
  62                 return ret;
  63
  64         host_kvm.mm_ops = (struct kvm_pgtable_mm_ops) {
  65                 .zalloc_pages_exact = host_s2_zalloc_pages_exact,
  66                 .zalloc_page = host_s2_zalloc_page,
  67                 .phys_to_virt = hyp_phys_to_virt,
  68                 .virt_to_phys = hyp_virt_to_phys,
  69                 .page_count = hyp_page_count,
  70                 .get_page = hyp_get_page,
  71                 .put_page = hyp_put_page,
  72         };
  73
  74         return 0;
  75 }
  76
  77 static void prepare_host_vtcr(void)
  78 {
  79         u32 parange, phys_shift;
  80
  81         /* The host stage 2 is id-mapped, so use parange for T0SZ */
  82         parange = kvm_get_parange(id_aa64mmfr0_el1_sys_val);
  83         phys_shift = id_aa64mmfr0_parange_to_phys_shift(parange);
  84
  85         host_kvm.arch.vtcr = kvm_get_vtcr(id_aa64mmfr0_el1_sys_val,
  86                                           id_aa64mmfr1_el1_sys_val, phys_shift);
  87 }
  88
  89 int kvm_host_prepare_stage2(void *mem_pgt_pool, void *dev_pgt_pool)
  90 {
  91         struct kvm_s2_mmu *mmu = &host_kvm.arch.mmu;
  92         int ret;
  93
  94         prepare_host_vtcr();
  95         hyp_spin_lock_init(&host_kvm.lock);
  96
  97         ret = prepare_s2_pools(mem_pgt_pool, dev_pgt_pool);
  98         if (ret)
  99                 return ret;
 100
 101         ret = kvm_pgtable_stage2_init_flags(&host_kvm.pgt, &host_kvm.arch,
 102                                             &host_kvm.mm_ops, KVM_HOST_S2_FLAGS);
 103         if (ret)
 104                 return ret;
 105
 106         mmu->pgd_phys = __hyp_pa(host_kvm.pgt.pgd);
 107         mmu->arch = &host_kvm.arch;
 108         mmu->pgt = &host_kvm.pgt;
 109         mmu->vmid.vmid_gen = 0;
 110         mmu->vmid.vmid = 0;
 111
 112         return 0;
 113 }
 114
 115 int __pkvm_prot_finalize(void)
 116 {
 117         struct kvm_s2_mmu *mmu = &host_kvm.arch.mmu;
 118         struct kvm_nvhe_init_params *params = this_cpu_ptr(&kvm_init_params);
 119
 120         params->vttbr = kvm_get_vttbr(mmu);
 121         params->vtcr = host_kvm.arch.vtcr;
 122         params->hcr_el2 |= HCR_VM;
 123         kvm_flush_dcache_to_poc(params, sizeof(*params));
 124
 125         write_sysreg(params->hcr_el2, hcr_el2);
 126         __load_stage2(&host_kvm.arch.mmu, host_kvm.arch.vtcr);
 127
 128         /*
 129          * Make sure to have an ISB before the TLB maintenance below but only
 130          * when __load_stage2() doesn't include one already.
 131          */
 132         asm(ALTERNATIVE("isb", "nop", ARM64_WORKAROUND_SPECULATIVE_AT));
 133
 134         /* Invalidate stale HCR bits that may be cached in TLBs */
 135         __tlbi(vmalls12e1);
 136         dsb(nsh);
 137         isb();
 138
 139         return 0;
 140 }
 141
 142 static int host_stage2_unmap_dev_all(void)
 143 {
 144         struct kvm_pgtable *pgt = &host_kvm.pgt;
 145         struct memblock_region *reg;
 146         u64 addr = 0;
 147         int i, ret;
 148
 149         /* Unmap all non-memory regions to recycle the pages */
 150         for (i = 0; i < hyp_memblock_nr; i++, addr = reg->base + reg->size) {
 151                 reg = &hyp_memory[i];
 152                 ret = kvm_pgtable_stage2_unmap(pgt, addr, reg->base - addr);
 153                 if (ret)
 154                         return ret;
 155         }
 156         return kvm_pgtable_stage2_unmap(pgt, addr, BIT(pgt->ia_bits) - addr);
 157 }
 158
 159 static bool find_mem_range(phys_addr_t addr, struct kvm_mem_range *range)
 160 {
 161         int cur, left = 0, right = hyp_memblock_nr;
 162         struct memblock_region *reg;
 163         phys_addr_t end;
 164
 165         range->start = 0;
 166         range->end = ULONG_MAX;
 167
 168         /* The list of memblock regions is sorted, binary search it */
 169         while (left < right) {
 170                 cur = (left + right) >> 1;
 171                 reg = &hyp_memory[cur];
 172                 end = reg->base + reg->size;
 173                 if (addr < reg->base) {
 174                         right = cur;
 175                         range->end = reg->base;
 176                 } else if (addr >= end) {
 177                         left = cur + 1;
 178                         range->start = end;
 179                 } else {
 180                         range->start = reg->base;
 181                         range->end = end;
 182                         return true;
 183                 }
 184         }
 185
 186         return false;
 187 }
 188
 189 static bool range_is_memory(u64 start, u64 end)
 190 {
 191         struct kvm_mem_range r1, r2;
 192
 193         if (!find_mem_range(start, &r1) || !find_mem_range(end, &r2))
 194                 return false;
 195         if (r1.start != r2.start)
 196                 return false;
 197
 198         return true;
 199 }
 200
 201 static inline int __host_stage2_idmap(u64 start, u64 end,
 202                                       enum kvm_pgtable_prot prot,
 203                                       struct hyp_pool *pool)
 204 {
 205         return kvm_pgtable_stage2_map(&host_kvm.pgt, start, end - start, start,
 206                                       prot, pool);
 207 }
 208
 209 static int host_stage2_idmap(u64 addr)
 210 {
 211         enum kvm_pgtable_prot prot = KVM_PGTABLE_PROT_R | KVM_PGTABLE_PROT_W;
 212         struct kvm_mem_range range;
 213         bool is_memory = find_mem_range(addr, &range);
 214         struct hyp_pool *pool = is_memory ? &host_s2_mem : &host_s2_dev;
 215         int ret;
 216
 217         if (is_memory)
 218                 prot |= KVM_PGTABLE_PROT_X;
 219
 220         hyp_spin_lock(&host_kvm.lock);
 221         ret = kvm_pgtable_stage2_find_range(&host_kvm.pgt, addr, prot, &range);
 222         if (ret)
 223                 goto unlock;
 224
 225         ret = __host_stage2_idmap(range.start, range.end, prot, pool);
 226         if (is_memory || ret != -ENOMEM)
 227                 goto unlock;
 228
 229         /*
 230          * host_s2_mem has been provided with enough pages to cover all of
 231          * memory with page granularity, so we should never hit the ENOMEM case.
 232          * However, it is difficult to know how much of the MMIO range we will
 233          * need to cover upfront, so we may need to 'recycle' the pages if we
 234          * run out.
 235          */
 236         ret = host_stage2_unmap_dev_all();
 237         if (ret)
 238                 goto unlock;
 239
 240         ret = __host_stage2_idmap(range.start, range.end, prot, pool);
 241
 242 unlock:
 243         hyp_spin_unlock(&host_kvm.lock);
 244
 245         return ret;
 246 }
 247
 248 int __pkvm_mark_hyp(phys_addr_t start, phys_addr_t end)
 249 {
 250         int ret;
 251
 252         /*
 253          * host_stage2_unmap_dev_all() currently relies on MMIO mappings being
 254          * non-persistent, so don't allow changing page ownership in MMIO range.
 255          */
 256         if (!range_is_memory(start, end))
 257                 return -EINVAL;
 258
 259         hyp_spin_lock(&host_kvm.lock);
 260         ret = kvm_pgtable_stage2_set_owner(&host_kvm.pgt, start, end - start,
 261                                            &host_s2_mem, pkvm_hyp_id);
 262         hyp_spin_unlock(&host_kvm.lock);
 263
 264         return ret != -EAGAIN ? ret : 0;
 265 }
 266
 267 void handle_host_mem_abort(struct kvm_cpu_context *host_ctxt)
 268 {
 269         struct kvm_vcpu_fault_info fault;
 270         u64 esr, addr;
 271         int ret = 0;
 272
 273         esr = read_sysreg_el2(SYS_ESR);
 274         BUG_ON(!__get_fault_info(esr, &fault));
 275
 276         addr = (fault.hpfar_el2 & HPFAR_MASK) << 8;
 277         ret = host_stage2_idmap(addr);
 278         BUG_ON(ret && ret != -EAGAIN);
 279 }